Barnaby Carr

Retinopathy of prematurity (ROP) screening has a staffing problem that predates any conversation about artificial intelligence. A 2006 survey of paediatric ophthalmologists and retinal specialists found that only about half were willing to manage ROP at all, and roughly one in five of those already doing it were considering stopping, citing falling reimbursement and the weight of medicolegal exposure (1). Two decades on, the pressure has not eased. Reviews of the U.S. workforce describe a shortage that is “serious and growing” even as improved neonatal survival pushes more extremely preterm infants into the screening pipeline (2). By one 2022 estimate, four U.S. states and 90% of counties had no paediatric ophthalmologist at all (3).

The reasons clinicians give for avoiding ROP are consistent across the literature: geographic dispersion of at-risk infants, the demand for weekly, year-round coverage, limited exposure to ROP during residency and fellowship training, and, above all, the disproportionate legal risk carried by a disease in which a missed diagnosis can mean permanent blindness in an infant (4,5). One review of fellowship training even found that a substantial share of clinical ROP exams are performed by trainees without direct attending supervision, a workaround for scarcity that itself raises the same liability question one layer down (6).

Telemedicine screening, in which trained non-physician imagers capture retinal photographs for remote grading, has become the standard response and the evidence for it is genuinely strong with multiple programs report the ability to detect treatment-warranted disease at sensitivities matching or approaching in-person examination (7). But telemedicine mainly solves distribution, it gets images out of remote units and onto a specialist’s screen. It does less to solve the underlying scarcity of specialists willing to grade those images and accept the liability that comes with the call.

This is the opening into which artificial intelligence has stepped. Machine learning models trained to classify ROP severity from retinal images have started to show real promise, including in low-resource settings. One study using smartphone-collected fundus videos found an algorithm classified ROP presence with greater sensitivity than a panel of three paediatric ophthalmologists, though the ophthalmologists retained better specificity (8). The pitch is straightforward – free up scarce specialist time, standardise a diagnosis that is already known to suffer from inter-grader disagreement, and extend screening to places that have never had reliable access to it.

What AI does not obviously solve is the liability question that drove clinicians away from ROP screening in the first place, it may just relocate it. Under current malpractice frameworks, courts generally still ask what the treating physician did, not whether the software failed. The “reasonable physician under similar circumstances” standard applies whether or not an algorithm was in the loop, meaning a doctor who follows a flawed AI recommendation without independent scrutiny can still be found solely liable (9). A 2023 systematic review of the medico-legal literature on AI diagnostic tools concluded that the existing regulatory framework for assigning liability is inadequate and unsettled, particularly where training data may be unrepresentative of the population the tool is later used on (10). Legal commentary on AI-assisted diagnosis more broadly describes liability as potentially shared among the treating physician, the hospital that deployed the tool, and the developer that built it, with the actual split depending on facts that have barely begun to be tested in court – whether the algorithm was defective, whether it was properly vetted before deployment and whether clinical staff were adequately trained to recognise its limits (11), (12).

For a subspecialty that already lost a generation of willing screeners to the fear of being solely blamed for a missed diagnosis, an unresolved question of shared liability is factor that needs addressing. A clinician deciding whether to grade ROP images today already weighs the catastrophic cost of a miss against relatively low reimbursement. Adding an algorithm to that workflow before liability law has caught up does not remove that calculation since it adds a second actor to the room without clarifying who answers for it.

None of this argues against AI-assisted ROP screening, but the workforce numbers make some form of automation increasingly hard to avoid. Treating algorithmic grading as a workforce fix while leaving its liability status ambiguous risks reproducing the exact dynamic that hollowed out the human workforce, just with a less accountable party standing where the ophthalmologist used to be.

References

1. Survey of pediatric ophthalmologists and retinal specialists on willingness to manage ROP (American Academy of Ophthalmology, 2006), as reported in: Screening and treatments using telemedicine in retinopathy of prematurity. PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC5398746/
2. Barrero-Castillero A, Corwin BK, VanderVeen DK, Wang JC. Workforce Shortage for Retinopathy of Prematurity Care and Emerging Role of Telehealth and Artificial Intelligence. Pediatr Clin North Am. 2020;67(4):725-733.
3. AI-Enabled Screening for Retinopathy of Prematurity in Low-Resource Settings. PMC. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12042057/
4. The Ophthalmologist. Pediatric Focus: Retinopathy of prematurity. 2024. https://theophthalmologist.com/issues/2024/articles/jun/pediatric-focus-retinopathy-of-prematurity
5. Can we reduce the burden of the current UK guidelines for retinopathy of prematurity screening? Eye. 2017. https://www.nature.com/articles/eye2017163
6. Training fellows for retinopathy of prematurity care: A Web-based survey. PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC3338950/
7. Screening and treatments using telemedicine in retinopathy of prematurity. PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC5398746/
8. AI-Enabled Screening for Retinopathy of Prematurity in Low-Resource Settings. PMC. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12042057/
9. Fault lines in healthcare AI – Part two: Who’s responsible when AI gets it wrong? Johns Hopkins Carey Business School. https://carey.jhu.edu/news/fault-lines-health-care-ai-part-two-whos-responsible-when-ai-gets-it-wrong
10. Cestonaro C, Delicati A, Marcante B, Caenazzo L, Tozzo P. Defining medical liability when artificial intelligence is applied on diagnostic algorithms: a systematic review. Front Med. 2023. doi:10.3389/fmed.2023.1305756
11. What Are the Legal Implications of AI and Machine Learning in Medical Malpractice Cases? PBG Law. https://www.pbglaw.com/blog/what-are-the-legal-implications-of-ai-and-machine-learning-in-medical-malpractice-cases/
12. When AI Gets the Diagnosis Wrong: Understanding Liability in Algorithm-Assisted Medical Errors. Davis & Davis. https://www.davis-davislaw.com/blog/ai-misdiagnosis-liability-medical-errors/